|
|
|
|
Xiao-Jing Wang
Perceptual Decision-Making by Slow Reverberation
In a Local Cortical Network
Brandeis University
Neurons at early sensory stages show very fast responses to inputs, which is presumably desirable for rapid information processing. Almost instantaneous response to a suddenly appearing input can be reproduced and understood by most existing cortical network models, even with biologically plausible membrane and synaptic time constants. On the other hand, In association cortical areas, higher-level computations (such as perceptual decision-making) take time (up to seconds). To understand the neural basis of higher cognitive processes, it is important to eludicate the unique cellular and network operations of association cortices. Here, I propose a class of local cortical network models characterized by slow temporal integration. I show that such a network provides a candidate circuit mechanism for perceptual decision-making in behaving monkeys.
I will focus on visual motion discrimination experiments of Newsome and colleagues, who discovered single neuron activities correlated with the animal's decision-making process. It is believed that decision is formed in the association cortices. Kim and Shadlen (Nature Neurosci. 2, 176-185 (1999)) showed that the prefrontal neuronal response during the stimulation, and persistent activity during the delay period, predict the monkey's judgment of motion direction. The biophysical circuit mechanisms underlying a decision-making cortical network remain unknown. I have investigated a cortical network model of perceptual decision making. The model assumes strong recurrent excitation, which is dominated by the NMDA receptors and is balanced by feedback inhibition. The findings include: (1) The slow NMDA receptor mediated reverberation provides a mechanism to temporally integrate the sensory evidence for motion direction, up to a few seconds. (2) Recurrent inhibition leads to competition between neural assemblies receiving conflicting input signals. (3) In reaction time simulations, decision is made when the firing activity of one of the neural assemblies reaches a fixed threshold. The model reproduces the experimental neurometric function and the reaction time data as function of the motion coherence level. (4) In delayed discrimination simulations, with sufficiently strong recurrent connections the network displays attractor dynamics of stimulus-selective persistent activity. This attractor network makes a categorical (binary) judgment derived from analog information about the stimulus. The experimental neurometric function can again be reproduced. Moreover, the network maintains the active memory of the decision across a delay period, in the form of persistent neural activity.
|
|
|
Saturday, December 7, 2024
|
|
The Swartz Foundation is on Twitter: SwartzCompNeuro
|
|
|
|